Research map / Probabilistic and scientific modeling / Dynamical systems

Dynamical modeling of nonlinear latent factors in multiscale neural activity with real-time inference

NeurIPS 2025 · Probabilistic and scientific modeling › Dynamical systems · approach: Low-rank and diffusion neural dynamics · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MRINE, a nonlinear multiscale dynamical model that fuses Poisson spike and Gaussian LFP signals with different timescales and missing samples, supporting real-time recursive decoding via modality-specific LDMs and a multiscale LDM backbone.

Problem

Real-time decoding of target variables from multiple neural modalities (spikes and LFPs) that have different sampling rates, different distributions, and missing samples; existing nonlinear multimodal dynamical models assume a shared timescale and/or use non-causal inference.

Approach

A multiscale encoder that first learns within-modality LDMs (Kalman-filtered, which handle missing samples by predicting forward) on MLP-embedded modality observations, then nonlinearly fuses the filtered embeddings with an MLP; a multiscale LDM backbone then gives causal Kalman filtering, with Poisson/Gaussian decoders and smoothness regularization on learned distributions.

Key results

• On the NHP grid reaching dataset with 5 spike and 5 LFP channels, MRINE reached 0.487 decoding accuracy versus 0.450 for the next-best baseline mmPLRNN.
• On the NHP center-out reaching dataset with 20 spike and 20 LFP channels, MRINE reached 0.649 versus 0.601 for MMGPVAE.
• Adding 20 LFP channels to 5 spike channels improved grid-reaching decoding accuracy by 14.7%, and adding 20 spike channels to 5 LFP channels improved it by 49.8%.

Measured effects reported in the paper

Surprising findings

• Zero-imputing missing LFP samples, a common practice, degraded performance of baselines such as mmPLRNN and MMGPVAE, while MRINE's degradation with differing timescales was smaller.

Limitations

• Assumes time-invariant multiscale dynamics, which may not hold for non-stationary data.
• Models may need intermittent retraining across days or sessions.
• Requires careful tuning of multiple smoothness regularization scaling hyperparameters.

Experimental setup

Datasets: stochastic Lorenz attractor simulations NHP grid reaching spike-LFP dataset NHP center-out reaching spike-LFP dataset neuropixel and calcium imaging visual cortex dataset

Benchmarks: NHP grid reaching NHP center-out reaching visual cortex frame ID decoding

Metrics: correlation coefficient R^2 latent reconstruction correlation

Baselines: MVAE MSID mmPLRNN MMGPVAE CEBRA

Builds on: DFINE mmPLRNN MSID MMGPVAE MVAE

Closest papers by meaning

• Integrating Multimodal Data for Joint Generative Modeling of Complex Dynamics (ICML 2024 · similarity 0.89)
• Probabilistic Decomposed Linear Dynamical Systems for Robust Discovery of Latent Neural Dynamics (NeurIPS 2024 · similarity 0.89)
• Multi-modal Gaussian Process Variational Autoencoders for Neural and Behavioral Data (ICLR 2024 · similarity 0.88)
• Meta-Dynamical State Space Models for Integrative Neural Data Analysis (ICLR 2025 · similarity 0.88)
• A scalable generative model for dynamical system reconstruction from neuroimaging data (NeurIPS 2024 · similarity 0.88)
• A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.87)